Intel Performance Skills — AI Agent Skills

A collection of AI agent skills for CPU performance analysis and optimization on x86 Linux. The skills give an agent structured, multi-step workflows for profiling with Linux perf, identifying performance anti-patterns in source code and profiling output, and optimizing benchmarks from the Phoronix Test Suite.

Example prompts

Once installed, you can talk to the agent naturally. A few examples:

Profile a workload and find hotspots

Please profile ./myapp --bench and tell me where the time goes.

Identify why a loop is slow

This function runs at 0.3 IPC despite low cache misses. What's wrong, and how do I fix it?

Optimize from source code alone

Review this dot-product loop for performance issues:

for (int i = 0; i < n; i++) sum += a[i] * b[i];

Cache-line contention

My server gets slower as I add threads. Can you use perf c2c to find false sharing?

Core-count scaling

This workload doesn't scale past 8 cores. Help me find the bottleneck.

Phoronix benchmark optimization

Profile pts/compress-zstd and optimize it. Report the before/after improvement.

Write new SIMD code correctly

Write a vectorized AVX2 float array scale function with CPU dispatch.

---

Contents

Directory	Purpose
skills/linux-perf/	Data collection skill: perf workflows, building blocks, hotspot reporting
skills/performance-patterns/	Pattern detection and fix playbooks: source code and profiling signals
skills/phoronix-test-suite/	Supporting skill: install, run, and optimize PTS benchmarks
skills/linux-tracing/	Off-CPU / I/O / syscall / latency tracing: bcc + bpftrace first, perf-tools fallback

---

Skills

linux-perf — profiling data collection

Profile Linux workloads with perf and identify performance bottlenecks.

When the user asks why code is slow, this skill guides data collection from hardware counters through to annotated hotspot reports. It provides:

• Flow 0 — perf stat on-CPU pre-check: confirm the work is CPU-bound before optimizing; pivots to linux-tracing if not.
• Flow T — toplev Top-Down (TMA): measured bottleneck category (Frontend/Backend/Bad-Spec/Retiring), routes to the right flow (Intel).
• Flow A — perf stat: quick IPC, cache-miss rate, and branch-miss rate in seconds.
• Flow B — perf record + perf report: which functions and source lines consume CPU.
• Flow C — perf c2c: cache-line contention, false sharing, and HITM events in multi-threaded code.
• Flow D — dual-profile scaling analysis: identify bottlenecks that only appear at high core counts.
• Flow E — structured hotspot report: formatted deliverable with top-function table, annotated source, and pattern observations.
• Annotate pattern scan: scan perf annotate output for scalar FP, narrow SIMD, serial accumulators, horizontal reductions, lock CAS, and memory pressure patterns.

When a pattern is identified, linux-perf hands off to performance-patterns for the fix.

Trigger phrases: "profile", "hotspot", "IPC", "cache miss", "false sharing", "HITM", "does not scale", "why is this slow", "where does time go".

---

performance-patterns — pattern detection and fixes

Detect and fix well-known performance anti-patterns — from source code or profiling output.

This skill owns the full catalog of patterns with detection signals and fix playbooks. It works standalone (source review only) or in combination with linux-perf (profiling data). Patterns covered:

Pattern	Detection signal	Fix
Serial accumulator	Low IPC + low cache misses; vaddss/vfmadd dominates annotate	Multiple independent accumulators + SIMD
Narrow SIMD	xmm/scalar in hot loop on AVX2/AVX-512 CPU	Vector width upconversion (zipper algorithm)
Missing vzeroupper	other_assists.avx_to_sse > 0; extreme cycle count on first SSE instruction	Insert vzeroupper before SSE transitions
Missing restrict	Aliasing preamble before vectorizable loop; two loop versions in annotate	Add restrict to pointer parameters
Test-and-Set spinlock	lock cmpxchg cluster in annotate; throughput drops with more threads	Test-and-Test-and-Set (TTAS)
False sharing	perf c2c HITM > 5%; different offsets written by different threads	Struct padding and layout reorganisation
Shared statistics counter	lock add/lock inc in hot path; true sharing in perf c2c	Per-CPU bucket array or local batching
Cache blocking	TMA DRAM-bound, low MLP / high miss latency	Loop tiling + stride-1 access / SoA
Memory bandwidth	TMA DRAM-bound, uncore bandwidth saturated	Non-temporal stores, prefetch, fewer streams
Frontend code layout	TMA Frontend-bound (DSB / i-cache / MS)	Cold-path splitting, PGO, LTO, reorder
Branch elimination	High branch-miss rate; unpredictable data-dependent branch	Branchless mask / cmov / SIMD blend

Also provides guidelines for writing new performance-sensitive C/C++ and SIMD code correctly from the start.

Trigger phrases: "optimize", "vectorize", "why is this slow", "review for performance", "improve throughput", "write a fast…", any _mm* intrinsics or inline asm with xmm/ymm/zmm registers.

---

phoronix-test-suite — supporting skill

Install, run, and optimize benchmarks from the Phoronix Test Suite.

Invoked automatically by linux-perf when the profiling target is a pts/<name> benchmark. Handles the full lifecycle: install, source extraction, rebuild with debug symbols, binary deployment, and result recording. Knows the correct install.sh build flags and source layout for each test.

Trigger: any pts/<name> reference, or the words "phoronix" / "phoronix-test-suite".

---

linux-tracing — off-CPU / I/O / syscalls / latency

Find where time goes when the bottleneck is not on-CPU.

linux-perf answers where the CPU spends cycles; linux-tracing answers the prior question — is the work even on-CPU? — and, when it isn't, traces off-CPU/blocked time, scheduler (run-queue) delay, disk/file I/O latency, page-cache misses, syscall storms, and tail / p99 outliers. It is eBPF-first: bcc ready-made tools (offcputime, runqlat, biolatency, syscount, …) and bpftrace one-liners, with Brendan Gregg's perf-tools (ftrace) as the no-eBPF fallback. Live tracing needs root (sudo protocol).

linux-perf Flow 0 (on-CPU pre-check) hands off to this skill when a workload turns out to be blocked rather than CPU-bound.

Trigger phrases: "off-CPU", "I/O wait", "disk latency", "syscalls", "scheduler delay", "context switches", "tail latency", "p99", "fast on average but slow sometimes".

---

Bundled tools

The skills invoke two bundled upstream tool repos as external programs — never copying their source (both are GPL-2.0, while these skills are MIT):

• pmu-tools/ (toplev, ocperf, ucevent) — powers linux-perf Flow T (TMA) and the Intel-event building blocks. The first run may need python3 pmu-tools/event_download.py (cached afterward). Core TMA needs perf_event_paranoid <= 1; uncore needs root.
• perf-tools/ (Brendan Gregg's ftrace scripts) — the no-eBPF fallback for linux-tracing. On modern kernels linux-tracing prefers system bcc (sudo apt-get install bpfcc-tools) and bpftrace.

---

Installation

This skill collection follows the open Agent Skills standard. Each skill directory contains a SKILL.md file that any compatible agent loads on demand. The same directories work across GitHub Copilot (CLI and VS Code), Claude Code, OpenAI Codex, Gemini CLI, and other compatible agents.

Agent	Project-level path	User-level path
GitHub Copilot CLI / VS Code	.github/skills/	~/.copilot/skills/
Claude Code	.claude/skills/	~/.claude/skills/
OpenAI Codex	.agents/skills/	~/.agents/skills/
Gemini CLI	.gemini/skills/	—
OpenCode	.opencode/skills/	~/.config/opencode/skills/

GitHub CLI (gh skill) — recommended

The easiest way to install across any supported agent. Requires GitHub CLI v2.90.0 or later.

# Install all skills (user-level)
gh skill install intel/intel-performance-skills linux-perf
gh skill install intel/intel-performance-skills performance-patterns
gh skill install intel/intel-performance-skills phoronix-test-suite
gh skill install intel/intel-performance-skills linux-tracing

Keep them up to date:

gh skill update linux-perf
gh skill update performance-patterns
gh skill update phoronix-test-suite
gh skill update linux-tracing

GitHub Copilot CLI

Skills are installed per-user under ~/.copilot/skills/:

cp -r skills/linux-perf ~/.copilot/skills/
cp -r skills/performance-patterns ~/.copilot/skills/
cp -r skills/phoronix-test-suite ~/.copilot/skills/
cp -r skills/linux-tracing ~/.copilot/skills/

GitHub Copilot in VS Code

GitHub Copilot in VS Code loads skills from .github/skills/ at project level or ~/.copilot/skills/ at user level. Project-level skills can be committed so the whole team benefits automatically:

# Project-level (commit to your repository)
mkdir -p .github/skills
cp -r skills/linux-perf skills/performance-patterns skills/phoronix-test-suite skills/linux-tracing \
    .github/skills/

# User-level (available in every project)
cp -r skills/linux-perf skills/performance-patterns skills/phoronix-test-suite skills/linux-tracing \
    ~/.copilot/skills/

VS Code also scans .claude/skills/ and .agents/skills/, so a single project-level installation covers GitHub Copilot, Claude Code, and Codex simultaneously.

Claude Code

Claude Code discovers skills in .claude/skills/ (project) or ~/.claude/skills/ (user):

# Project-level
mkdir -p .claude/skills
cp -r skills/linux-perf skills/performance-patterns skills/phoronix-test-suite skills/linux-tracing \
    .claude/skills/

# User-level
cp -r skills/linux-perf skills/performance-patterns skills/phoronix-test-suite skills/linux-tracing \
    ~/.claude/skills/

OpenAI Codex

Codex reads skills from .agents/skills/ at project level and ~/.agents/skills/ at user level:

# Project-level
mkdir -p .agents/skills
cp -r skills/linux-perf skills/performance-patterns skills/phoronix-test-suite skills/linux-tracing \
    .agents/skills/

# User-level
cp -r skills/linux-perf skills/performance-patterns skills/phoronix-test-suite skills/linux-tracing \
    ~/.agents/skills/

Gemini CLI

Gemini CLI reads project skills from .gemini/skills/:

mkdir -p .gemini/skills
cp -r skills/linux-perf skills/performance-patterns skills/phoronix-test-suite skills/linux-tracing \
    .gemini/skills/

OpenCode

OpenCode has native skill support and reads skills from .opencode/skills/:

mkdir -p .opencode/skills
cp -r skills/linux-perf skills/performance-patterns skills/phoronix-test-suite skills/linux-tracing \
    .opencode/skills/

For user-level installation, copy to ~/.config/opencode/skills/ instead.

---

License

Copyright (C) 2026 Intel Corporation. Released under the MIT License.